News · 2026-09-04

AISLE found six curl CVEs after frontier-model scans found none

AISLE says its AI-assisted vulnerability-discovery system found six new curl vulnerabilities that were fixed in curl 8.22.0, after cited frontier-model scanners returned no findings. Curl's own security advisories confirm the six disclosures, making this a real AI-cyber result rather than an unreviewed benchmark claim. The significance is not that a model magically broke curl; it is that a specialized search, reproduction, and maintainer-review pipeline found accepted bugs in a mature codebase.

Key facts

• AISLE announced the finding wave on September 2, 2026; curl's release and advisory tables list six associated low-severity CVEs in curl 8.22.0.
• The reported issues include an OpenSSL-provider use-after-free, pinning bypass, native CA-store reuse flaw, cookie parsing flaw, wolfSSL ordering issue, and public-suffix-list cookie-scope failure.
• All six were reported by Stanislav Fort, according to curl's security table; one, CVE-2026-82208, does not affect the curl command-line client.
• Primary source: AISLE's disclosure, corroborated by curl's release table.

The six bugs are deliberately unglamorous, which is one reason the story is credible. They are narrow state and validation failures, not a breathless claim that AI found a universal memory-corruption catastrophe. The detailed curl advisories describe backend-specific behavior around certificate providers and trust stores, while other entries cover cookie parsing and public-suffix handling. This is the kind of code where a mistake can survive for years because it occurs only in a particular configuration, ordering, or character-handling path.

That specificity also explains why models alone are not the story. Finding a potential flaw in a large C codebase is only the first step. A useful system must trace the state machine, construct a reproduction, determine which builds are affected, distinguish a bug from intended behavior, eliminate duplicates, explain impact, and survive scrutiny by the maintainers who will have to fix it. Think of a metal detector on a beach: a louder detector finds more signals, but the valuable system is the one that separates coins from bottle caps, maps the location, and gives the owner enough evidence to dig.

AISLE says frontier scanners had come back empty before its run. The appropriate reading is careful. This was not necessarily a head-to-head contest of the same base model with the same time budget. AISLE's claimed edge is a pipeline around the model—search, verification, and workflow—not a public proof that one hidden neural network is categorically more capable than another. Its June curl post reported another six CVEs fixed in 8.21.0, which makes the September result more than a one-off, but it is still evidence from a single unusually hardened project.

Daniel Stenberg, curl's maintainer, supplies an important independent calibration. In a May post, he wrote that a Mythos scan had yielded one low-severity curl CVE and that, for curl, the hype felt mostly like marketing—while adding that modern AI analyzers were materially useful. In late August he posted the public tally “Mythos: 0 / Aisle: 29.” AISLE reproduces Linux maintainer Greg Kroah-Hartman's response: “I'm seeing the same for Linux as well. No idea what Aisle is doing differently, but wow...” Because the quote appears in AISLE's post, it should be understood as AISLE's reproduction of his comment, not an independently hosted statement here.

The broader security lesson is that AI can raise the throughput of adversarial code review when it is attached to a disciplined evidence loop. This is closer to a skilled security team with better search and tireless test generation than to fully autonomous offensive hacking. The accepted-CVE gate is crucial: it protects against the false-positive problem that makes many automated vulnerability tools expensive to use. It also protects against inflated claims of impact.

There is a more uncomfortable implication. The same technical ingredients—code navigation, hypothesis generation, test construction, and validation—can aid offense as well as defense. That is why the surrounding controls matter: authorization, isolated test environments, logging, disclosure processes, and human review. Jailbreaking and red teaming are relevant concepts, but this incident is about software analysis rather than attacking a model's guardrails.

The honest caveat is severity and scope. All six disclosed issues are low severity, many affect only particular TLS backends or platforms, and no public evidence here establishes a general win rate across projects. A responsible next test would use several codebases, pre-registered evaluation criteria, matched budgets, blinded maintainer triage, and published false-positive rates. Until then, the solid claim is still notable: a specialized AI-assisted pipeline produced six real, maintainer-accepted security fixes in curl where other cited scans did not.

Primary source, verified: read the paper →

Ground Truth. (2026, September 4). AISLE found six curl CVEs after frontier-model scans found none. Ground Truth. https://groundtruth.day/news/aisle-found-six-curl-cves-after-frontier-scanners-found-none.html

@misc{groundtruth:aisle-found-six-curl-cves-after-frontier-scanners-found-none,
  title  = {AISLE found six curl CVEs after frontier-model scans found none},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/news/aisle-found-six-curl-cves-after-frontier-scanners-found-none.html}
}

Topics: cybersecurity · ai-security · vulnerabilities · code-security · curl · agents · red-teaming